BRUMO25

New methods optimize LLM inference by analyzing confidence dynamics

Two new research papers propose methods to optimize the inference time of large language models by analyzing their confidence levels during reasoning. The first paper, EAGer, uses token-wise entropy to dynamically alloc…